Human action recognition in video by 'meaningful' poses

TLDR: A graph theoretic technique for recognizing actions at a distance by modeling the visual senses associated with human poses, and the results clearly demonstrate the superiority of the approach over the present state-of-the-art.

Abstract: We propose a graph theoretic technique for recognizing actions at a distance by modeling the visual senses associated with human poses. Identifying the intended meaning of poses is a challenging task because of their variability and such variations in poses lead to visual sense ambiguity. Our methodology follows a bag-of-words approach. Here "word" refers to the pose descriptor of the human figure corresponding to a single video frame and a "document" corresponds to the entire video of a particular action. From a large vocabulary of poses we prune out ambiguous poses and extract 'meaningful' [6] poses - for each action type in a supervised fashion - using centrality measure of graph connectivity [16]. The number of 'meaningful' poses per action is determined by setting a bound on the centrality measure. We evaluate our methodology on four standard activity recognition datasets and the results clearly demonstrate the superiority of our approach over the present state-of-the-art.